Before01
Company overview
- Counting done by eye
- Errors from stacking
- Missing count records
AI
Vision inspection
Zero-defect quality management completed by precise, vision-AI-based inspection
OVERVIEW
What the eye misses,a trained model sees the same way every time
From image capture through pre-processing and feature extraction to judgment and feedback, as one flow.
01Image capture
Cameras and lighting photograph each part under the same conditions every time.
Noise removal and normalization
02Image pre-processing
Noise is removed and brightness and size are normalized, so the image is ready for analysis.
AI pattern analysis
03Feature extraction and recognition
A trained AI model finds the features a judgment needs, such as shape and defect location.
Pass/fail judgment
04Judgment and feedback
Each part is judged good or defective, and the result is fed back into process improvement.
Feedback to image capture
KEY FEATURES
Four capabilities that replace visual inspection
From surface inspection through defect detection and dimensional measurement to quality statistics.
Capability Index04
Vision Capability
01 / 04
01
Automated visual inspection
Defect detection
Precision dimension measurement
Quality statistics and process optimization
It instantly detects minute defects with a high-performance vision system and deep-learning algorithms. By establishing a uniform quality standard with no variation between operators, it raises shipment reliability.
Management Scope
- High-speed image processing
- Full-inspection automation
- Multi-angle camera control
- Real-time judgment monitoring
- Lighting and exposure optimization
- Unstructured data collection
Real-time image capture
High-resolution cameras and optics capture surface images during production in real time
WORKFLOW
The workflow behind each capability
For each capability, the order it runs in and the check it passes before moving on.
Recognition performance starts with a consistent image
- 01
Capture design
camera · lighting · distance
- 02
Timing linkage
speed · sensor trigger
- 03
Image correction
brightness · resolution · region
- 04
Quality check
blur · occlusion · exposure
Recorded alongside
01
Capture position/
02
Trigger/
03
Raw/
Operating check
Is the inspection target clearly visible?
YES
Valid image → model judgment
NO
Occluded/blurred → recapture · review
An example workflow · the detailed links and criteria are set to suit your floor.
SYSTEM ARCHITECTURE
What it connects,and the work it leads to
From the data that goes in to the results on the floor.
01 / INPUT
01Images · videoCamera · lighting · capture trigger
02Item standardsShape · spec · inspection area
03Judgment historyOK/NG labels · LOT
CONNECTED OPERATIONS
AI Vision
Image correction · synchronization
Object, defect and count recognition
Judgment · exceptions · history
03 / OUTPUT
✓Inspection resultsPosition · type · size · count
✓Review screenOriginal · marked area · judgment
✓Production linkLOT · output · defect analysis
01 / INPUT
01Images · videoCamera · lighting · capture trigger
02Item standardsShape · spec · inspection area
03Judgment historyOK/NG labels · LOT
CONNECTED OPERATIONS
AI Vision
Image correction · synchronization
Object, defect and count recognition
Judgment · exceptions · history
CONNECTED OPERATIONS
AI Vision
Image correction · synchronization
Object, defect and count recognition
Judgment · exceptions · history
03 / OUTPUT
✓Inspection resultsPosition · type · size · count
✓Review screenOriginal · marked area · judgment
✓Production linkLOT · output · defect analysis
Linked by common identifiers, master data and historyResults checked → fed into the next run
Who uses itQuality inspectorsProduction engineersShipping and logistics staff
METHOD
How it is built
Floor imagery gathered and learned, then returned as a verdict.
Collection01 →
Data Ingestion
Line-scan camera
3D vision sensor
Thermal camera
Raw imagesDefect annotationsPass/fail historyTrigger signal
Image preprocessing
Lighting correction · resolution normalization
Trigger sync
Tied to conveyor speed · precise capture timing
Target System
Image storage
Vector DB
MetadataDefect history
Modeling & analysis02 →
AI/ML Modeling
Analysis & prediction
Object detection
YOLO v10, 11
Defect segmentation
Scratches · pinholes · cracks
Defect classification
CNN, Transformer
3D measurement
Shape deviation · dimensions
Feedback & refinement
Model evaluation
Action history
MLOps retraining
Service03
Operational visibility
Dashboard
KPI report
Process root cause
Reinforcement learning (RLHF)
Intelligent maintenance
Failure prediction
Instant alerts past thresholds / automatic equipment stop
Advanced deep-learning technology overcomes the limits of visual inspection.
Electronic components
Robot vision
Automotive
Precision machining
Secondary batteries
Food/pharmaceuticals
Electronic components
Robot vision
Automotive
Precision machining
Secondary batteries
Food/pharmaceuticals
A camera judges every piece, including the flaws a person misses.
- Inspection images
- Lighting and capture conditions
- Pass criteria
- Defect type history
- Judgment feedback
AI Vision
- Defect candidate detection
- Type and grade classification
- Criteria retraining
- Detection rate
- Missed defects
- False rejects
- Inspection effort
EFFECT
Business impact
What it leaves behind — in detection rate, misses, false rejects and inspection effort.
- 01
Higher detection rate
High-resolution imaging and deep learning catch the fine defects the eye misses, and hold one standard with no drift between judgments.
Detection ratejudgment driftdefect size - 02
Fewer misses
Every unit is inspected, so nothing slips through a sample, and a model per defect type learns the patterns that are easiest to miss.
Miss ratefull inspectiondefect types - 03
Fewer false rejects
The boundary that used to call a good part bad is re-set from data, so sound product stops being scrapped or re-run.
False rejectsthresholdsrework - 04
Less inspection effort
Inspection staff move to higher-value work, and the time spent re-setting the line for a different item comes down.
Staffingchangeoverhourly output
CASE STUDY
Case study
Manual counting automated, taking accuracy to around 99.9%.
Customer industry
Primary steel manufacturing
After02
Implementation
- Cognex camera capture
- Job event triggers
- Busy-scene detection
Business impact
Inspection automation where AI vision does the counting